COMPARISON OF AERIAL PHOTOGRAPHS AND LASER SCANNING DATA AS
METHODS FOR OBTAINING 3D FOREST STAND PARAMETERS
M. Schardt a, W. Hruby a, M. a Hirschmugl a, R. Wack b , M. Franke b
a
Graz University of Technology -mathias.schardt@tugraz.at, wolfi1@sbox.tugraz.at
manuela.hirschmugl@joanneum.at
b
Institute of Digital Image Processing, Joanneum Research Graz - (roland.wack,martina.franke)@joanneum.at
KEY WORDS: Photogrammetry, Forestry, Automation, Matching, DEM/DTM, Laser scanning
ABSTRACT:
The aim of this study is an evaluation of the use of (semi-) automatic software solutions for the derivation of crown surface models
and further forest information like vegetation height and vertical stand structure. It gives an overview of the up-to-date reachable
accuracies of crown surface models and vegetation heights from traditional aerial images in comparison with the results from laser
scanner data. Concerning the surface models, two algorithms (feature-based matching approach - ISAE and an area-based matching
approach - RSG) were tested. The results were evaluated with three different methods. Despite careful investigations, the evaluations
didn’t lead to a definite preference of one of the algorithms. On visual interpretation, the ISAE seemed to be too smooth, while the
RSG method produced higher errors within the single tree height measurements. For the area based comparison, both led to quite the
same accuracies. Concerning the calculation of the vegetation height, the second main part of the study was the evaluation of the
available terrain model. As the laser scanner terrain model could be used as ground truth, the difference is depicted through
statistical figures. It turned out, that the official terrain model with errors of more than 10 m is not appropriate for measuring
vegetation height. It can be summarised, that both photogrammetric algorithms lead to erroneous results in forest stands with a
strong, but small-patched vertical structure, but to good surface models in quite homogeneous stands. In combination with an
accurate terrain model – for example derived by laser scanning - the mean stand height for these parts can be calculated with an
appropriate accuracy.
1. INTRODUCTION
Aerial images have long been successfully used for measuring
various forest inventory and environmental parameters. Forest
parameters that can be derived from aerial photos are for
example tree species, tree age, crown closure and forest health.
By means of photogrammetric methods additionally threedimensional information such as vertical stand structures can be
obtained.
It is obvious, however, that not all important forest parameters
can be derived from aerial images. Significant shortcomings
have been identified in mapping the terrain surface
characteristics and terrain roughness beneath vegetation stands,
which can provide the basis for assessing tree or stand height,
for the planning, construction and optimisation of forest road
networks or for the assessment of avalanche risk in alpine
protection forests. Another problem is that automatic
photogrammetric methods available on the market are not
appropriate for the derivation of detailed tree wise information
or to measure vertical structures within the vegetation stand.
This is due to the fact that parts of the tree crowns are
shadowed and the “visible” shape of a tree is varying due to its
geometrical position on the stereo images.
The parameters ”terrain roughness” and ”vertical stand
structure” can basically be derived from airborne laser scanner
data as not all laser pulses will be reflected on the vegetation
surface, but a significant proportion will penetrate into the
interior of the vegetation stand and will be reflected by the
ground surface and by vegetation layers within the stand.
All papers should be sent to the meeting organisers in digital
form. However, if in exceptional circumstances, the paper
cannot be prepared digitally, it must be prepared on A4 paper
according to these guidelines, and sent to the organisers for
scanning.
2. OBJECTIVES
In the paper the potential of aerial photos and laser scanner data
for the assessment of vertical forest stand parameters such as
terrain roughness and tree/stand height will be discussed.
Additionally digital terrain models of forested areas resulting
from laser scanner data will be compared with digital terrain
models of the Austrian Federal Mapping Agency. For the
processing of the aerial photos the software packages Intergraph
(ISAE) and RSG (Remote Sensing Software Package Graz)
were used. The processing of laser scanning data was
performed with the Software IMPACT. Both, IMPACT and
RSG were developed at the Institute of Digital Image
Processing of Joanneum Research. The results presented are
derived from the project “Derivation of Forest Parameters by
Means of High Resolution Remote Sensing Data” which was
financed by the “Styrian Woodcluster” (Hirschmugl et al.
2004).
3. DATA AND TEST SITE
The investigation was performed on the base of laser scanner
data and aerial photos covering our test site “Burgau” located in
the eastern part of Styria. The test site is characterized by hilly
and flat terrain. The main tree species are beech (fagus
sylvestica), oak (quercus sp.), spruce (picea abies), pine (pinus
sylvestris) and larch (larix decidua). The relief varies from flat
to hilly. The forests are mainly owned by farmers and thus play
an important economic role in this region.
The laser scanner campaign was carried out in the year 1999
using the TopoSys scanner because of its high measurement
- 272 -
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVI - 8/W2
density, steep viewing angle and capability of providing both
first and last pulse modes. The data captured during this
campaign comprises:
First pulse: 23.8.1999; flight height 800m; 4-5
points per m2
Last pulse: 26.3.1999; flight height 800m; 4-5
points per m2
Infrared aerial photos were taken of the same area in the year
1998 at a scale of approximately 1:15.000 to 1:17.000.
Additionally, a digital terrain model (DTM) of this area with a
horizontal resolution of 10 m was purchased from the Federal
Calibration and Survey Office (Bundesamt für Eich- und
Vermessungswesen = BEV) in Vienna. The Z-accuracy (height)
of this DTM is officially specified between ±5 m and ±20 m for
forested areas (BEV).
4. DIGITAL TERRAIN AND SURFACE MODELS
DERIVED FROM LASER SCANNER DATA
In this chapter a brief description of the method used for the
derivation of DTMs from laser scanner data will be given. The
laser DTM will be used for the quality assessment of the DTMs
available at the Federal Austrian Mapping Agency. More
detailed information on the filter method is given in Ruppert et
al., 2000; Wack et al. 2002 and Ziegler et al. 2001.
Raw last pulse data were processed in a regular 1 x 1m grid
using a multi-resolution method in combination with an
improvement by the integration of gradients. The method
searches in a first step the raw data for the minimum values
within the respective resolution and increases the minimum by
an appointed amount as a function of the gradient in the
neighborhood. Smoothing and thresholding algorithms are
performed in each resolution step, thus resulting in
corresponding pixel values. The method starts with a coarse
resolution (e.g. 10 m) and compares the results with better
resolutions (e.g. 7, 5, 3, 1 m) step by step. Hence, the method
allows forest floor information to be obtained even in dense
stands, as it is assumed that within 10 x 10 m at least one laser
signal comes from the ground. The multi-resolution method decisions whether to take the coarse or better resolution are
made by thresholding - assures that ground results are taken
from better resolutions whenever there is a ground signal, in
other cases (dense stands) the value from the coarse resolution
is taken. The accuracy obtained by more than hundred
terrestrial GPS and geodetic measurements varies between 18
and 45 cm. The high accuracy of the DTM´s derived from laser
scanner are precise enough to use it as “ground truth” for the
quality check of the relatively rough DTM´s of the Austrian
Mapping Agency.
5. PROCESS FOR GENERATING A SURFACE MODEL
FROM STEREO AERIAL PHOTOGRAPHS
For the purpose of constructing crown surface models from
stereo aerial photographs, the software packages Intergraph
(ISAE) available at the Institute for Remote Sensing and
Photogrammetry of Technical University Graz and the software
RSG (Remote Sensing Software Package Graz) developed at
the Institute for Digital Image Processing of the Joanneum
Research were tested. The description of the methods focusses
on the matching processes and 3D reconstruction available in
the two software packages since it is primarily these image
processing steps that determine the accuracy of the resulting
surface models.
Process steps preceding the matching, carried out with the
Intergraph software, were ‘image orientation’ as well as a rough
registration of image pairs for minimising parallaxes. This latter
process step limits the search area for image correlation and
increases the obtainable matching accuracy.
5.1 Building a surface model with ISAE (feature based
matching)
For calculating the elevation model, the ISAE process employs
a “feature based algorithm” which is used step by step within
the pyramid structure (“coarse-to-fine strategy”), (Kraus, 2000).
On each level of the pyramid homologous pixels are extracted.
This is used to form 3D intersections in the object space and
generate a DTM from bilinear finite elements. ISEA projects
individual mesh elements of a baseline DTM (the baseline is a
horizontal plane through a point in the centre of the area to be
processed) into the image and thereby generates the
homologous image areas.
It means that for each image a list of possible intersections is
generated. Then these lists of characteristics are compared with
each other and a search is carried out for suitable pairs of points
for assigning to the image. For the similarity measurements, the
ISAE uses the interest value and the correlation coefficient
calculated with the help of the Förstner operator (Förstner,
1986). This process is carried out for each of the image levels
within the image pyramid. The points thus found are entered
into the next finer level of the pyramid. At the end a ‘least
squares matching’ is carried out on the finest pyramid level in
order to further improve the accuracy of the image correlation.
In detailed tests it was possible to show that the highest
accuracies are achieved with the following parameter settings:
Epipolar line distance: for this purpose the parameter
‘epipolar line distance’ (ELD) indicates in which epipolar lines
‘feature points’ are searched for in the image, and it therefore
influences the number of points. The best results have been
obtained with an ELD of 1.
Parallax bound.: indicates a limit value in the object space in
which measured 3D points are included in the calculation
(maximum height difference between individual points). The
best results were obtained with values around 20.
Size of correlation window: 5 x 5 pixel
Number of pyramids: 10
Grid width X/Y: defines the step intervals of the DTM grid in
metres. The best values were obtained with a step interval of 2
metres, which is for the present data the smallest achievable
step interval in Intergraph.
Smoothing weight: defines the degree to which the terrain
model is smoothed in the reconstruction process. Since for the
purpose of this study it was important to maintain ‘sharp edges’,
which were the result of e.g. gaps between adjacent trees,
clearings or edges of the forest, this parameter was given no
weighting (smoothing weight 1).
5.2 Reconstructing a surface model with RSG(area based
matching)
A further surface model was calculated at the Institute for
Digital Image Processing of the Joanneum Research Society in
- 273 -
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVI - 8/W2
Graz using the RSG software (RSG field guide) and using a
hierarchical correlation algorithm which was also based on
pyramid levels. This was a plane-based approach and the
correlation method selected was cross correlation.. Detailed
Information can be found in (RSG Feld Guide) and (Franke,
2004)
In order to optimise the matching results, all results with a back
matching distance of larger than 3 were filtered out. In a further
step, a weighted grid interpolation (weighted mean) of the
irregularly distributed points cloud was carried out. Below we
list the parameters with which the best results were obtained.
Number of pyramides: 7
Size of reference window: 7 x 7 pixel
Size of search window: 121(col) x 41(line)
Grid width: pixelsize of the scanned image
Figure 2. Profile of: heterogenous conifer stand, thinning and
mean timber
6. COMPARISON
In this section we will, on the one hand, compare the crown
surface models obtained with the laser scanner data with those
obtained with stereo aerial photographs. In addition, we will
investigate whether it is possible to determine the height of
trees or the vertical forest stand structure from a combination of
the terrain models of the Federal Calibration and Survey Office
(BEV), which are available on a nation-wide level, with the
crown surface models derived from the stereo aerial
photographs. Since for this purpose the quality of the BEV
models is an important prerequisite, these models are compared
with the digital terrain model obtained from the laser scanner
data.
6.1 Comparsion of crown surface models
The crown surface models (RSG, ISAE and laser scanner
models) are visually compared on the basis of profiles that
represent the different situations in the test site Burgau:
Figure 3. Profile of: forest border
The profile lines in profile 1 (figure 1) show that both models
provide a good representation of the average stand height in a
dense homogenous stand. The ISAE model is a bit more
smoothed. Of course, compared to the laser scanner data, it is
not possible to ‘see through’ to the forest floor. However, when
using detailed topographical models, both approaches provide
good results for determining the average stand height.
Figure 1. Profile of homogenous conifer stand, mean/strong
timber
The profile lines in profile 2 (figure 2) show that it is normally
not possible to record individual trees (such as e.g. single trees)
with either the RSG or the ISAE method. If in some instances
this can be achieved, the value between the two standards is
subject to interpolation, which leads to an inaccuracy for the
value of the height in between (see figure 2 between a distance
of 20 and 40 m). This results in inaccuracies of 10 m and more
in both methods and the surface models have in these cases only
limited value.
The profile lines in profile 3 (figure 3) show that the edges of
forests are displayed very differently by the different models.
Naturally, the laser scanner model displays the optimum
boundary by showing a steep edge on the edge of the forest. By
contrast, the other two models ‘fudge’ the edge to varying
degrees and thus lead to inaccuracies in the vegetation height
both in the forest, where the values are too low, as well as
outside the forest where the values are too high.
- 274 -
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVI - 8/W2
The results of the visual comparison of the profile lines can be
summarised as follows: the vertical structures of individual
trees are well depicted in the laser scanner data but are only
scantly shown in the crown surface models based on the stereo
aerial photographs. This is particularly the case where there are
differences in height with only small gaps or taking up only
small areas, such as e.g. in stands with great variations in
vertical structure, where there are boundary areas between
stands of different ages, or at the edges of forests. In addition, it
became apparent (see e. g. figure 1 between 70 and 90 m, figure
3 at 20 m distance) that this “smoothing effect” was more
pronounced in the crown surface models created with the ISAE
software. This is due to the smaller grid width of the RSG
model, so the stand structure can be mapped with more detail.
A quantitative evaluation of the different methods for the
creation of crown surface models was carried out, firstly, by an
area-based comparison of absolute heights in different forest
stands and secondly, by comparing the height of 43 reference
trees that were measured in the terrain. The results of the areabased comparison are shown in table 1, the results of the
comparison with the reference trees in table 2.
Type of forest stand
Young homogenous
conifer stand (approx. 85
x 100 m)
Heterogenous multilayer conifer stand
(approx. 150 x 150 m)
Single-layer dense
deciduous/conifer stand
(approx. 140 x 150 m)
Old structurally
differentiated conifer
stand with clearings
(approx. 165 x 100 m)
Old deciduous stand
without clearings
(approx. 70 x 60 m)
Average
difference
between RSG
and LS surface
models [cm]
290
Average
difference
between ISAE
and LS surface
models [cm]
289
500
493
263
267
771
745
Table 2 shows the comparison between the terrestrially
measured trees and the vegetation heights from the different
calculations. The terrain model used for the derivation of the
vegetation height was in both cases the one based on the last
pulses from the laser scanner.
The average error from the RSG-based calculation is slightly
more than 8m and hence a bit greater than the ISAE-based
calculation with approximately 7.5m. Therefore this
comparison indicates that the ISEA software is to be preferred.
6.2 Comparsion of terrain models
The determination of vegetation height, which is one of the
relevant parameters from the forestry point of view, requires not
only a crown surface model but also a terrain model. Since we
cannot assume that laser scanner data are available for all areas,
it is intended to check the available BEV model to see whether
it is suitable for this purpose. Therefore as already mentioned
the laser scanner model was used as ground truth. So, the BEVterrain model was subtracted from the laser scanner terrain
model. The following table (Table 3) shows the statistical
parameters of the resulting difference image. That means the
maximum differences are 34 m and approx. 9 m respectively. In
general, the BEV terrain model tends to result in too high
values.
Minimum
[cm]
-3400
Maximum
[cm]
881
Mean
value [cm]
1033
Standard
deviation [cm]
860
Table 3: Statistical parameters between the laser scanner floor
model and the BEV terrain model
206
207
Table 1: Average differences in height (unit values) in selected
stands between the RSG and ISAE models
respectively and the heights obtained with laser
scanning
Amount of average error
[cm]
imaging of the crown surface. The results also show that the
deviations are less pronounced in the case of more homogenous
conifer forests and deciduous forests with less differentiations
in the vertical surface structure compared to those of stands
with strong differentiations in the vertical surface structure. Old
stands with lesser density also show high errors, but this might
be only because of the higher amount of terrain points mapped
by the laser scanner.
RSG
817
With an average deviation of over 10 m the quality check
corresponds to the BEV statements. This means that this terrain
model is not suitable for the determination of the vegetation
height since the errors are to large. It also means that when
using photogrammetric methods for forestry purposes it is
necessary to employ an accurate terrain model such as can be
created from laser scanner data.
7. SUMMARY
ISAE
743
Table 2: Average error in the determination of the height of 43
reference trees in crown surface models created with
the photogrammetric methods (RSG and ISAE)
compared with terrestrial measurements.
The results listed in table 1 show that the height values of the
photogrammetrically derived height models are at variance with
the height values of the laser scan models. Of course, these
deviations are due, on the one hand, to the laser scanner values
penetrating the crowns and scanning the floor. On the other
hand, these deviations are also the errors inherent in the
In this study we have tested two different algorithms for
creating crown surface models: RSG and ISAE. Both were
subjected to a quality assessment, comparing them with laser
scanner data and terrestrial surveys of trees and also carrying
out a visual interpretation. When visually assessing the profiles
created for the two methods, the ISAE provides a somewhat
more smoothed surface model compared to the RSG. This has a
negative effect, particularly in areas covering the edge of the
forest.
With respect to the quantitative error in establishing the
vegetation height of more structured crown surface, both
algorithms led to similar results, when comparing with
terrestrially surveyed reference trees, the average error of ISAE
- 275 -
International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. XXXVI - 8/W2
is smaller. Both models are generally not capable of recording
and displaying differences in level that extend only over a small
area and are therefore not suitable for measuring the height of
individual trees. However, in combination with an accurate
terrain model the average height of forest stands can be
determined very satisfactorily. In the comparison of the terrain
models we were able to confirm that, under forest stands, the
BEV model produced errors of over 10 m and is therefore not
suitable for the determination of the height of forest stands.
8. LITERATURE
References from Journals:
Förstner, W. (1986): A Feature Based Correspondence
Algorithm for Image Matching, IAPRS, Vol. 26(3), pp. 150 166.
Ruppert, G.; Wimmer, A.; Beichel, R.; Ziegler, M.: ”An
adaptive multi-resolutional algorithm for high precision forest
floor DTM generation”, Proceedings of AeroSense’2000, Laser
Radar Technology and Applications V, 4035, April 2000.
Schardt, M.; Ziegler, M.; Wimmer A.; Wack R. & Hyyppä J.
(2002): Assessment of Forest Parameters by Means of Laser
Scanning
Proceedings of the ISPRS Commission III
Symposium Graz, pp. 302-309.
Wack, R. & Wimmer, A. ( 2002): Digital Terrain Models from
Airborne Laser scanner Data – a Grid based approach
Proceedings of the ISPRS Commission III Symposium Graz,
pp. 293-296.
Ziegler, M.; Wimmer, A. & Wack, R. (2001): DTM generation
by means of airborne laser scanner data – an advanced method
for forested areas. Proceedings of the 5th Conference on
Optical 3-D Measurement Techniques Vienna, pp. 97 - 102.
References from Other Literature:
Franke, M. (2003): Generierung digitaler Geländemodelle aus
hochauflösenden optischen Satellitenbilddaten, Unpubl. Master
Thesis, Technical University of Graz, Graz, 114 p.
Hirschmugl et. al. (2004): Machbarkeitskonzept zur
großflächigen Implementierung von Informationssystemen auf
GIS-Basis in Waldverbänden, Unpubl. Project Report,
Technical University of Graz, Graz, 72 p.
References from Books:
Kraus, K. (2000): Photogrammetrie, Band 3 – Topografische
Informationssysteme, Dümmler-Verlag.
References from websites:
BEV - Bundesamt für Eich- und Vermessungswesen :
www.bev.gv.at
RSG Field Guide, JOANNEUM RESEARCH, Institut für Digitale
Bildverarbeitung, Wastiangasse 6, 8010 Graz, available at:
http://dib.joanneum.ac.at/rsg
- 276 -